From: Peng Hao <[email protected]>
Windows I/O, such as the real-time clock. The address register (port
0x70 in the RTC case) can use coalesced I/O, cutting the number of
userspace exits by half when reading or writing the RTC.
Guest access rtc like this: write register index to 0x70, then write or
read data from 0x71. writing 0x70 port is just as index and do nothing
else. So we can use coalesced mmio to handle this scene to reduce VM-EXIT
time.
In our environment, 12 windows guests running on a Skylake server:
Before patch:
IO Port Access Samples Samples% Time% Avg time
0x70:POUT 20675 46.04% 92.72% 67.15us ( +- 7.93% )
After patch:
IO Port Access Samples Samples% Time% Avg time
0x70:POUT 17509 45.42% 42.08% 6.37us ( +- 20.37% )
Cc: Paolo Bonzini <[email protected]>
Cc: Radim Krčmář <[email protected]>
Cc: Eduardo Habkost <[email protected]>
Cc: Peng Hao <[email protected]>
Signed-off-by: Peng Hao <[email protected]>
Signed-off-by: Wanpeng Li <[email protected]>
---
v1 -> v2:
* add the original author
Documentation/virtual/kvm/00-INDEX | 2 ++
Documentation/virtual/kvm/api.txt | 7 +++++++
Documentation/virtual/kvm/coalesced-io.txt | 17 +++++++++++++++++
include/uapi/linux/kvm.h | 5 +++--
virt/kvm/coalesced_mmio.c | 16 +++++++++++++---
virt/kvm/kvm_main.c | 2 ++
6 files changed, 44 insertions(+), 5 deletions(-)
create mode 100644 Documentation/virtual/kvm/coalesced-io.txt
diff --git a/Documentation/virtual/kvm/00-INDEX b/Documentation/virtual/kvm/00-INDEX
index 3492458..4160620 100644
--- a/Documentation/virtual/kvm/00-INDEX
+++ b/Documentation/virtual/kvm/00-INDEX
@@ -9,6 +9,8 @@ arm
- internal ABI between the kernel and HYP (for arm/arm64)
cpuid.txt
- KVM-specific cpuid leaves (x86).
+coalesced-io.txt
+ - Coalesced MMIO and coalesced PIO.
devices/
- KVM_CAP_DEVICE_CTRL userspace API.
halt-polling.txt
diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt
index d10944e..4190796 100644
--- a/Documentation/virtual/kvm/api.txt
+++ b/Documentation/virtual/kvm/api.txt
@@ -4618,3 +4618,10 @@ This capability indicates that KVM supports paravirtualized Hyper-V TLB Flush
hypercalls:
HvFlushVirtualAddressSpace, HvFlushVirtualAddressSpaceEx,
HvFlushVirtualAddressList, HvFlushVirtualAddressListEx.
+
+8.19 KVM_CAP_COALESCED_PIO
+
+Architectures: x86, s390, ppc, arm64
+
+This Capability indicates that kvm supports writing to a coalesced-pio region
+is not reported to userspace until the next non-coalesced pio is issued.
diff --git a/Documentation/virtual/kvm/coalesced-io.txt b/Documentation/virtual/kvm/coalesced-io.txt
new file mode 100644
index 0000000..4a96eaf
--- /dev/null
+++ b/Documentation/virtual/kvm/coalesced-io.txt
@@ -0,0 +1,17 @@
+----
+Coalesced MMIO and coalesced PIO can be used to optimize writes to
+simple device registers. Writes to a coalesced-I/O region are not
+reported to userspace until the next non-coalesced I/O is issued,
+in a similar fashion to write combining hardware. In KVM, coalesced
+writes are handled in the kernel without exits to userspace, and
+are thus several times faster.
+
+Examples of devices that can benefit from coalesced I/O include:
+
+- devices whose memory is accessed with many consecutive writes, for
+ example the EGA/VGA video RAM.
+
+- windows I/O, such as the real-time clock. The address register (port
+ 0x70 in the RTC case) can use coalesced I/O, cutting the number of
+ userspace exits by half when reading or writing the RTC.
+----
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index b6270a3..9cc56d3 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -420,13 +420,13 @@ struct kvm_run {
struct kvm_coalesced_mmio_zone {
__u64 addr;
__u32 size;
- __u32 pad;
+ __u32 pio;
};
struct kvm_coalesced_mmio {
__u64 phys_addr;
__u32 len;
- __u32 pad;
+ __u32 pio;
__u8 data[8];
};
@@ -949,6 +949,7 @@ struct kvm_ppc_resize_hpt {
#define KVM_CAP_GET_MSR_FEATURES 153
#define KVM_CAP_HYPERV_EVENTFD 154
#define KVM_CAP_HYPERV_TLBFLUSH 155
+#define KVM_CAP_COALESCED_PIO 156
#ifdef KVM_CAP_IRQ_ROUTING
diff --git a/virt/kvm/coalesced_mmio.c b/virt/kvm/coalesced_mmio.c
index 9e65feb..fc66a834 100644
--- a/virt/kvm/coalesced_mmio.c
+++ b/virt/kvm/coalesced_mmio.c
@@ -83,6 +83,7 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
ring->coalesced_mmio[ring->last].phys_addr = addr;
ring->coalesced_mmio[ring->last].len = len;
memcpy(ring->coalesced_mmio[ring->last].data, val, len);
+ ring->coalesced_mmio[ring->last].pio = dev->zone.pio;
smp_wmb();
ring->last = (ring->last + 1) % KVM_COALESCED_MMIO_MAX;
spin_unlock(&dev->kvm->ring_lock);
@@ -149,8 +150,12 @@ int kvm_vm_ioctl_register_coalesced_mmio(struct kvm *kvm,
dev->zone = *zone;
mutex_lock(&kvm->slots_lock);
- ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, zone->addr,
- zone->size, &dev->dev);
+ if (zone->pio)
+ ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, zone->addr,
+ zone->size, &dev->dev);
+ else
+ ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, zone->addr,
+ zone->size, &dev->dev);
if (ret < 0)
goto out_free_dev;
list_add_tail(&dev->list, &kvm->coalesced_zones);
@@ -174,7 +179,12 @@ int kvm_vm_ioctl_unregister_coalesced_mmio(struct kvm *kvm,
list_for_each_entry_safe(dev, tmp, &kvm->coalesced_zones, list)
if (coalesced_mmio_in_range(dev, zone->addr, zone->size)) {
- kvm_io_bus_unregister_dev(kvm, KVM_MMIO_BUS, &dev->dev);
+ if (zone->pio)
+ kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,
+ &dev->dev);
+ else
+ kvm_io_bus_unregister_dev(kvm, KVM_MMIO_BUS,
+ &dev->dev);
kvm_iodevice_destructor(&dev->dev);
}
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 8b47507f..a587fb9 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -2936,6 +2936,8 @@ static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
#ifdef CONFIG_KVM_MMIO
case KVM_CAP_COALESCED_MMIO:
return KVM_COALESCED_MMIO_PAGE_OFFSET;
+ case KVM_CAP_COALESCED_PIO:
+ return 1;
#endif
#ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
case KVM_CAP_IRQ_ROUTING:
--
2.7.4
2018-07-12 09:59+0800, Wanpeng Li:
> From: Peng Hao <[email protected]>
>
> Windows I/O, such as the real-time clock. The address register (port
> 0x70 in the RTC case) can use coalesced I/O, cutting the number of
> userspace exits by half when reading or writing the RTC.
>
> Guest access rtc like this: write register index to 0x70, then write or
> read data from 0x71. writing 0x70 port is just as index and do nothing
> else. So we can use coalesced mmio to handle this scene to reduce VM-EXIT
> time.
>
> In our environment, 12 windows guests running on a Skylake server:
>
> Before patch:
>
> IO Port Access Samples Samples% Time% Avg time
>
> 0x70:POUT 20675 46.04% 92.72% 67.15us ( +- 7.93% )
>
> After patch:
>
> IO Port Access Samples Samples% Time% Avg time
>
> 0x70:POUT 17509 45.42% 42.08% 6.37us ( +- 20.37% )
>
> Cc: Paolo Bonzini <[email protected]>
> Cc: Radim Krčmář <[email protected]>
> Cc: Eduardo Habkost <[email protected]>
> Cc: Peng Hao <[email protected]>
> Signed-off-by: Peng Hao <[email protected]>
> Signed-off-by: Wanpeng Li <[email protected]>
> ---
> v1 -> v2:
> * add the original author
>
> Documentation/virtual/kvm/00-INDEX | 2 ++
> Documentation/virtual/kvm/api.txt | 7 +++++++
> Documentation/virtual/kvm/coalesced-io.txt | 17 +++++++++++++++++
> include/uapi/linux/kvm.h | 5 +++--
> virt/kvm/coalesced_mmio.c | 16 +++++++++++++---
> virt/kvm/kvm_main.c | 2 ++
> 6 files changed, 44 insertions(+), 5 deletions(-)
> create mode 100644 Documentation/virtual/kvm/coalesced-io.txt
>
> diff --git a/Documentation/virtual/kvm/00-INDEX b/Documentation/virtual/kvm/00-INDEX
> index 3492458..4160620 100644
> --- a/Documentation/virtual/kvm/00-INDEX
> +++ b/Documentation/virtual/kvm/00-INDEX
> @@ -9,6 +9,8 @@ arm
> - internal ABI between the kernel and HYP (for arm/arm64)
> cpuid.txt
> - KVM-specific cpuid leaves (x86).
> +coalesced-io.txt
> + - Coalesced MMIO and coalesced PIO.
> devices/
> - KVM_CAP_DEVICE_CTRL userspace API.
> halt-polling.txt
> diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt
> index d10944e..4190796 100644
> --- a/Documentation/virtual/kvm/api.txt
> +++ b/Documentation/virtual/kvm/api.txt
> @@ -4618,3 +4618,10 @@ This capability indicates that KVM supports paravirtualized Hyper-V TLB Flush
> hypercalls:
> HvFlushVirtualAddressSpace, HvFlushVirtualAddressSpaceEx,
> HvFlushVirtualAddressList, HvFlushVirtualAddressListEx.
> +
> +8.19 KVM_CAP_COALESCED_PIO
> +
> +Architectures: x86, s390, ppc, arm64
> +
> +This Capability indicates that kvm supports writing to a coalesced-pio region
> +is not reported to userspace until the next non-coalesced pio is issued.
> diff --git a/Documentation/virtual/kvm/coalesced-io.txt b/Documentation/virtual/kvm/coalesced-io.txt
> new file mode 100644
> index 0000000..4a96eaf
> --- /dev/null
> +++ b/Documentation/virtual/kvm/coalesced-io.txt
> @@ -0,0 +1,17 @@
> +----
> +Coalesced MMIO and coalesced PIO can be used to optimize writes to
> +simple device registers. Writes to a coalesced-I/O region are not
> +reported to userspace until the next non-coalesced I/O is issued,
> +in a similar fashion to write combining hardware. In KVM, coalesced
> +writes are handled in the kernel without exits to userspace, and
> +are thus several times faster.
> +
> +Examples of devices that can benefit from coalesced I/O include:
> +
> +- devices whose memory is accessed with many consecutive writes, for
> + example the EGA/VGA video RAM.
> +
> +- windows I/O, such as the real-time clock. The address register (port
> + 0x70 in the RTC case) can use coalesced I/O, cutting the number of
> + userspace exits by half when reading or writing the RTC.
> +----
> diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
> index b6270a3..9cc56d3 100644
> --- a/include/uapi/linux/kvm.h
> +++ b/include/uapi/linux/kvm.h
> @@ -420,13 +420,13 @@ struct kvm_run {
> struct kvm_coalesced_mmio_zone {
> __u64 addr;
> __u32 size;
> - __u32 pad;
> + __u32 pio;
Paolo, do you think we can rename the field without breaking userspace
builds?
> };
>
> struct kvm_coalesced_mmio {
> __u64 phys_addr;
> __u32 len;
> - __u32 pad;
> + __u32 pio;
> __u8 data[8];
> };
>
> diff --git a/virt/kvm/coalesced_mmio.c b/virt/kvm/coalesced_mmio.c
> @@ -149,8 +150,12 @@ int kvm_vm_ioctl_register_coalesced_mmio(struct kvm *kvm,
> dev->zone = *zone;
>
> mutex_lock(&kvm->slots_lock);
> - ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, zone->addr,
> - zone->size, &dev->dev);
> + if (zone->pio)
> + ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, zone->addr,
> + zone->size, &dev->dev);
> + else
> + ret = kvm_io_bus_register_dev(kvm, KVM_MMIO_BUS, zone->addr,
> + zone->size, &dev->dev);
This would be better readable as
ret = kvm_io_bus_register_dev(kvm, zone->pio ? KVM_PIO_BUS : KVM_MMIO_BUS,
zone->addr, zone->size, &dev->dev);
thanks.